- Inventory truth needs an explicit reservation and release policy.
- Channel events must be idempotent and reconciled, not trusted blindly.
- Peak readiness is a tested operating state, not a hope.
Define what available means
Available stock is not the same as on-hand stock. It is on-hand minus reservations, safety stock, allocations and unresolved exceptions. We define these components and the moment they change. Stores, warehouse, web and marketplaces may have different fulfilment rights, but they share the same underlying truth.
Reservations have an owner and an expiry. An abandoned checkout should release stock; a confirmed order should convert the reservation and create a fulfilment task. The policy is explicit, so support teams are not guessing why an item disappeared.
- On-hand, reserved, safety and allocated stock
- Reservation ownership and expiry
- Channel-specific fulfilment rights
- Clear conversion and release rules
Order events deterministically
Marketplace and storefront events arrive with different guarantees. We use idempotent consumers, stable event keys and a per-SKU sequence where the source provides one. Duplicate order creation is prevented at the command layer, not by hoping the queue delivers once.
Reconciliation compares channel orders, payment settlements and warehouse movements. Exceptions go to an operations queue with context. A daily spreadsheet reconciliation is a symptom that the event model is incomplete.
- Idempotent channel consumers
- Stable event and order identities
- Payment, order and stock reconciliation
- Exception queues with operational context
Rehearse the peak
Peak readiness covers more than load tests. We check cache behaviour, payment provider limits, queue back-pressure, fulfilment capacity and stock reconciliation under burst traffic. Kill switches and degraded modes are defined for non-critical features.
After the event, measure oversells, cancelled lines, payment failures, fulfilment lag and service contacts. Those metrics expose whether the architecture supports the commercial promise.
- Peak load and failure scenario tests
- Queue back-pressure and payment limits
- Degraded modes and kill switches
- Post-event oversell and lag metrics
Questions we hear in the room
Can marketplaces and physical stores share inventory?
Yes, with explicit fulfilment rights, reservation rules and reconciliation. Not every channel should be allowed to promise every unit.
Do we need a new ERP?
Not necessarily. The first step is usually a reliable inventory and event model, with the ERP integrated through clear contracts.
